Interpretation of land cover/land use and accuracy assessment

A pre-interpretation of the aerial photograph mosaic was made to get to know the image con-tent and stratify the area in zones to be visited (ALBERTZ 2009). 157 sites were selected for field investigation. They were visited between 2005 and 2006 (Figure III.4). The selection of the sites was based on the following criteria:

 Unknown surface patterns in the aerial photographs;

 Coverage of all zones and of majority of land cover/use types observed in the photos and

 Accessibility of the sites/terrain.

During field work the sites were characterized by their main land cover/land use type, floristic composition, percentage canopy cover, elevation, slope, aspect and former natural or human disturbances.

FIGURE III.4. Sampled sites for visual aerial photograph interpretation (Source: GPS field work;

Boundaries of ABNP from Law 64-00; Villages, roads from topographic maps 1:50,000, ICM 1983/1984 and field work; trails digitized on aerial photograph mosaic)

UTM WGS84 coordinates of each location were saved to a handheld Trimble GeoXM GPS, running TerraSync Standard Edition 2.50 with an accuracy of ± 5 m (Trimble Navigation Li-mited, Westminster/USA). Photos were taken of the sites. 82 of the collected GPS points were superimposed over the aerial photograph mosaic in ArcGIS 9.1. The patterns of the aerial photos were correlated to the ground observation to elaborate the classification sceme. Alto-gether 75 points were set aside for accuracy assessment.

The determination of discrete land cover/land use units is an important step of aerial photo-graph interpretation (CINGOLANI et al. 2004). Three criteria had to be fulfilled by the classifi-cation scheme: The classes of the system had to be in concordance with the objectives of the study and had to be identifiable in the aerial photographs (CINGOLANI et al. 2004). The classes of the system had to allow a division in more detailed subclasses so that these can be inte-grated for monitoring purposes on future aerial photograph series or large-scale satellite im-ages. The management direction of the national park had to participate and to agree on the classification system of Armando Bermúdez National Park. They were interested to differen-tiate between the following classes: Dense pine forests, open pine forests, humid broadleaf forests, cloud forests, mixed forests, Prestoea montana forests (manaclares), shrubland (ma-torral), land with ferns (Calimetal), coffee with shade, coffee without shade, cultivated land (conucos), grassland and sparse vegetation.

Analysis of the aerial photographs brought two results: a.) It was not possible to differentiate between all the classes due to the quality of the aerial photographs. b.) Some classes could have been delineated, but it would have been too time consuming (see Discussion).

The process of land cover/land use mapping was based on the characteristic elements of aerial photograph interpretation (RABBEN 1960; HILDEBRANDT 1996; HARALICK 1979; J EN-SEN 2007; LILLESAND et al. 2008; ALBERTZ 2009). These elements are: Location, shape, size, pattern, tone/color, texture, shadows, height/depth and site/situation/association. The location of an object is defined by its coordinates. Shape is related to the general form or outline of an object. Its size is scale dependent and can be measured accurately on an orthophotograph (length, width, perimeter, area, volume). Pattern stands for the spatial arrangement of objects in an aerial photograph and is characteristic for many features (JENSEN 2007) (here e.g. coffee plants on a plantation). Color depends on the proportions of energy reflected from the surface materials in the different electromagnetic spectra. Shades of color are better distinguished by the human eye than tones of gray (SABINS 1996).

The characteristic spatial interrelationships between tones in an image are referred to as tex-ture (HARALICK 1979). Texture is created by tonal repetitions of groups of objects that

indivi-dually would be too small to be discerned. Texture is extremely scale dependent (HARALICK

1979) and is a long recognized value in aerial photograph interpretation (ANDRESEN et al.

2002; LANGANKE et al. 2004). Pattern and texture are not affected by bidirectional reflectance, other than tone and shadow (HARALICK et al. 1973). Shadow helps to estimate the heights of objects, but can also impede the identification of others. Objects in a stereopair can be visua-lized in three dimensions and their heights measured (LILLESAND et al. 2008; ALBERTZ 2009).

Site, situation and association are related to each other and used synergistically together to conclude about the aerial photograph content (JENSEN 2007).

Taking into account these elements, the objectives of the study, the physiognomy and canopy structure of the vegetation in Armando Bermúdez National Park, 13 cover-classes could be differentiated: three related to the main class “Forest”, three to the main class “Rangeland”, two to the main class “Agricultural Land” and five to the main class “Additional Categories”.

The digitizing scale of 1:2,500 resulted appropriate to detect small fields, resulting in a mini-mum mapping unit of 4,500 m² (0.45 ha) for the main classes “Rangeland”, “Agricultural Land” and “Additional Categories”. The natural and semi-natural vegetation units in “Forest”

were digitized at a scale of 1:4,000. This working scale was a compromise between overall information extraction and man hours to conclude the work. Each class in the classification scheme was characterized by pattern, color and texture (see Results). Pine and broadleaf fo-rests were additionally described according to their crown physiognomy and forest structure (HUDSON 1991).

Surface patterns on the mosaic were reviewed in 3D on the analogue stereopairs with a lens stereoscope (one of the four stereoscopic viewing techniques described by JENSEN 2007).

Then homogeneous patches were digitized on the mosaic and labeled according to the proper-ties of the vegetation in the polygon. On-screen digitization is a straightforward method for obtaining spatial units of land cover/land use and accurate area measurements (JENSEN 2007).

Additionally the DTM, a hydrology shapefile and an aspect rasterfile derived from the DTM, were opened as collateral data.

Accuracy assessment of the interpretation was undertaken with the 75 field points set aside for accuracy assessment in a contingency table with rows representing the classified data and the columns the reference data (FOODY 2002). Statistical measures as the producer‟s accuracy, user‟s accuracy and the overall classification accuracy (percentage correctly classified) were calculated (CONGALTON 1991). The producer‟s accuracy describes the percentage of reference points of a certain class assigned to the same class in the image (error of omission). The user‟s accuracy indicates the percentage of pixels of a class in the classification in concordance with

the reference data (error of commission). Landscape metrics such as the percent area occupied by a land cover/land use type (P) and number of patches (NP) (just for the classes in “Agricul-tural Land”) were extracted per water basin and for the entire protected area in ArcGIS 9.1 (GARDNER et al. 1987).

Zonal statistics were calculated to relate the land cover/land use types to altitude and slope.

Nearest distances between the centerpoints of the polygons in “Agricultural Land” to the bor-der of the protected area and to the villages were calculated. The workflow of aerial photo-graph interpretation and analysis of land cover and land use classes is shown in Figure III.5.

Pre-interpretation of aerial photographs

↓ Field work

↓

Determination of classification scheme

↓

Visual interpretation and delineation of polygons

↓

Accuracy assessment for selected classes

↓

Calculation of landscape metrics/Statistics

↓ Final map result

FIGURE III.5. Workflow of visual interpretation of aerial photographs of Armando Bermúdez Na-tional Park (ALBERTZ 2009, modified)